 You see images all over the internet where people stick lines and sometimes they make the more elaborate lines and they track a whole field of players and they say, yeah, we can do biomechanics on a whole field now. If you only have joint centers. That's not mathematically observable. Okay. My name is Scott Selby, and I'd like to thank Stuart for giving me this opportunity to talk. I'm representing two companies on C motion and Thea. And one of the strange things about companies is sometimes people pick your topics for you and your titles. And pose estimation is a big topic, certainly far bigger than than one lecture. So this is my experience with pose estimation and I'm only going to talk about technologies that I've been involved with and algorithms I've been involved with. There are several categories that I want to talk about. But principally, I'm talking about biomechanics. So pose estimation as it's typically used in biomechanics. And mostly from optical motion capture although I'll cheat a little bit and present some x-ray stuff. What are we talking about with pose estimation and a video is kind of worth a thousand words. And in biomechanics, we tend to talk about the skeleton more than we talk about the surface of the body. And that's predominantly because the muscles move the bones and the muscles are controlled in some way by the brain. And if we want to understand what the brain is doing or we want to understand what's possible for a person. One point is to track the skeleton. Tracks in term I will introduce in a minute, but we want to get the movement of some model of the body. But we never get to measure that we get to measure the surface of the body, either a sensor on the body or just the image itself. We typically represent the surface as an inferred skeleton. So most of this talk will be talking about the skeleton. Here's a segment. Now, when we talk about pose estimation, we talked about biomechanics, the fundamental thing we talk about in kinematics is a segment. And as everybody can see, it's just a coordinate system. And people are expecting to see a bone that's decorated with some stuff. But in actual fact, it's just the reference frame, this coordinate system that is relevant. And we can give it some object, give it a cube here, just because it's easier to see. And the cube can move around in space. We typically assume that it's a rigid object so that when it moves, it maintains its shape. And to describe that movement, we describe a point on this rigid body. And since it's rigid, it really reflects all points on the body. And we describe it by a very simple looking equation that we can take a point in the coordinate system that moved. We can describe it in terms of a world coordinate system by one transformation. One four by four matrix that contains the orientation translation, and any kind of scaling or misshaping but typically we're referring only to orientation and translation. That's what I'm talking about a transformation. Well, it just so happens that in the C3D format the ubiquitous format for storing motion capture data and biomechanics are as referred to as a rotation group. There's a point group for points, an analog group for one dimensional signals, and this rotation group which represents a four by four signal. And actually pose estimation is the process of solving for our. This may seem like a little bit of a back step, but there's some terminology we had to get used to. We talk about salient features salient means important bits in an image on a film. And the features we tend to talk about are sometimes anatomical, but on the left here we're showing, you know, traditional marker based motion capture and the salient features are actually the markers themselves. But as far as the algorithms and the cameras go, there really isn't a body there are markers, it finds those features those features are the salient features, and we give them names. So we endow them with some purpose some feature by associating a name with them. And how we associate the names is idiosyncratic. There's no common way of doing it although there's a fantastic book here that's shown in the bottom left that does nothing but name anatomical anatomical parts of the of the body. And has this wonderful description of how to palpate for them how to find them how to count how to name them. And almost nobody uses the names. So we never really have a roadmap back to a really elaborate definition of a point. So, anyways we give them the features. When it comes to markerless tracking. So tracking from video cameras without features. The salient features are actually specific anatomical features. And this kind of makes more sense to call them salient features we're talking about a knee joint or a hip joint, I've got a very, very simplified old fashioned markerless tracking skeleton assigned to the subject here. And the process without going into detail is rather than just put a marker on and have a camera find a marker is that we teach a neural network to recognize the feature. There's a little panel in the bottom. You see you take an image can be the image of anyone doesn't matter what they're wearing. Mostly these are humans in the case of a human biomechanics I'm only talking about humans here. And we train a network to find features like the ones on the right. The number of features has a very similar meaning to the number of markers to put on a body, more features more markers, a more complex model. Sometimes features can just be a contour in the x-ray image or it could be actually a full silhouette. And we could try to find a full silhouette or part of silhouette, but I'll probably keep coming back to these things because there's something critically important about how markers behave. Okay, the starting point. I said the segment has just a coordinate system, but it's, it's not fair to say we can put it anywhere in the world, it wouldn't be reasonable. We could put it on an anatomical location that kind of makes some sense to us. And there are some roadmaps. The image on the left is from the Rizzoli marker set. And it defines names and defines locations and there is explicitly reference frames associated with that. So here's a simple one and then a complicated I really struggled with this slide. Here's a shank. And so we define the coordinate system typically aligned with an axis that passes from the distal to the proximal end of the segment. And for a segment like this shank, but seems pretty straightforward. For a segment like the pelvis is not exactly clear where the distal and proximal ends are so it's a bit more complicated. And to find that coordinate system based on some other anatomical features, which we use to align the coordinate system. And what I've done is this cluttered little thing is to show how many different features there are and what seems really obvious to us when people often say media lateral knee. They mean a whole lot of things. So we have to be very, very careful how we place them. I'm going to come back to that. So by putting markers on you could use some kind of a pointer. You could use an x-ray. So this is an EOS scanner. You could use a biplane video radiography to label it. These are all kind of complicated ways of doing the anatomy, but if you care about the anatomy, sometimes it's worth that amount of effort to take for accurate. So this is a study done with biplane or video radiography. And it's just one example of the challenge we face. Despite the rules we have for locating and what we're looking at here is a lateral view of the pelvis. We're looking at the acetabulum and we're saying, where's the hip. So the hip for marker based data is defined from the markers on the pelvis. We define from movement, we've got a number of different ways here of doing it. One of them, the green one happens to be from biplane or video radiography. And it's the one that tends to sit in the middle of the acetabulum, which is where you expect all the others very depending on the subject, very dependent on depending on the placement. So I have to be careful when we have a very precise definition that in practice, when you go to implement that by doing something with your hands. It's never that precise. Okay, so we have a reference frame, a reference frame, there's some markers. And these don't have to be the same. You can put a marker or you put a feature for marker list, anywhere you want on a segment. So we tend to choose ones that are convenient for us. Sometimes we use the ones that are we use for the reference frames because we don't want to put on 135 markers on a subject so when we want to double the load for some of them. There's two kinds of tracking that I talked about that have been involved with and then I'll get to a variation of them. The first one, euphemistically called six degree of freedom tracking, very old. This is from the 1970s literature really. And it says, we're going to track every segment independently of all other segments. So, regardless of how well or how badly we define other segments, at least this one segment we might get right. And so perhaps in our research, only one of them really, really matters to us. We want to treat it independently of all the others. So it's not tainted by any errors in the others. And then there's a very old least squares fit that says if I know where those markers are when the subject's not moving, I can find them when the subject is moving. And this often gets called segment optimization because we're having a best fit of markers into an expected location for those markers. Now, segments aren't typically in isolation in the body so there's a large group of research that spend time saying, but we know they're connected. We know the knee is connected or we know the hip is connected to the pelvis. Surely that information is relevant to what we're doing. And in an ideal world in which every segment was measured well, that'd be true. That'd be great. If we could define those constraints properly. For a control subject, a healthy subject, we're probably going to be pretty good. We're going to be doing some pathology, or perhaps an athlete that's just so extreme in terms of their flexibility, maybe not so good but nevertheless. This is referred to as multi body optimization because not one segment gets treated as many segments as there are linked together by your definition. They all have to satisfy this optimization process. There's a little figure there which is just showing you figures of coordinate systems, typically at the proximal end of a joint. Every segment has one parent segment. And this is gets confusing in the middle of the body because the feet are distal to use the anatomical description and hands are distal. That's true if they're all distal somewhere in the middle of the body that they meet. So somewhere in the middle of the body we don't actually have a constraint explicitly. Just kind of strange but there's usually some workarounds to get it. Now, here's the critical bit you're supposed to take away because this is what I spent much of my career dealing with mathematical observability. Do you have enough information in your salient features to uniquely identify the position and orientation of a segment for segment optimization is really obvious. You have to have at least three non collinear markers or features on a segment. That's it. You can have more. It is an optimization problem after all you could put 100 markers on a segment if you wanted to spend the time doing that. That's really easy. I'm not going to go into this in the lecture here. But let me say that if you want to know if your multi body pose is mathematically observable. It's not as easy as counting markers. You can always put three features on every segment and that would be satisfied. But often people are trying to use less markers. And the number of markers required depends on where they are. And depends on the pose of the body, mathematically. The objective function which you're minimizing to get pose the Jacobian of that for a given state for the markers that features you've been presented has to be full rank. Well, that's a mouthful. I'm not taking it apart in this amount of time. Let's say there is a mathematical way to determine it. But almost nobody does. I think most people cross their fingers. And hope that it's all okay. So what do you do about it? Well, you can just put lots of markers everywhere. There's a problem with that too. Not surprisingly. So this is what is a pretty straightforward common multi segment foot model markers put on this is the Rizzoli foot model. I'm not going to go into detail but if you look at that cloud of points beside it all the colored markers on that skeleton. That looks like way more than the markers you see in the picture. Well, what we've done is take three different multi segment foot models. People don't really agree. So if you take all of the interpretations of the marker set and just put them down and this is only three. You see points everywhere. Now, if you say if you tell us which markers that you use, well, that's fine. But I know so many people would say I want to use this marker set. Yeah, but you know my my patients walk on the outside of their foot so all the ones on the lateral side of the foot I'm going to take off. Is that okay? Well, you decide it's okay, but it's no longer coined by that original markers. This is one that is so common as this is the most trivial of the markerless solutions. You get a very limited number of joint centers. Is that observable? So this is what you see you see images all over the internet where people stick lines and sometimes they make the more elaborate lines and they track a whole field of players and they say, yeah, we can do biomechanics on a whole field now. But if you only have joint centers, that's not mathematically observable. So you have to realize that you kind of have a skeleton, but you don't have a reliable 3D position and orientation of all those segments. And almost regardless of the kind of constraints you put on, you never will. So if you're going to do it, if you can see all the blue dots kind of wavering here. If you're going to have a markerless solution or a marker based solution, you put lots and lots of markers on. Now with marker based solution, it takes time to put all those markers on. Markerless, it just takes time to train the algorithm. It's kind of you get given that for free. Okay, so tracking. We call pose estimation pose estimation actually refers to a single frame of data and the tracking refers to the collection of contiguous frames. Do we have some gold standard that says, you know, we did it well. We have some gold standard that we can compare to and say, are we accurate. Do we get the numbers we want. Well, this has been proposed as a gold standard, but you've seen that little picture is a very cluttered image because it is cluttered. That's the biplanar video radiography lab here at Queens University. And yeah, it's cluttered to get our anatomical reference frames, typically, you use some kind of imaging protocol. You don't try to place markers on the surface here, you actually look at the bones and say, are there features on the bones that we can define our reference frame for each one. Now, yes, it requires you to have a CT scanner and MR scan of the subject, bit of a nuisance, but you can get a pretty good bone. That's the representation of the subjects bones. Now in biplanar video video radiography hard to say biplanar refers to two images. And so we see two different planes, two different views. And we put together the 3d position of the bones. And before someone jumps all over me, you say, well that image based registration surely that's not a gold standard. You can verify your biplanar system. Well, you get some very accommodating subjects who allow you to implant beads in their bones, and then you track those beads. Those beads themselves then become your salient features and you solve the problem in a six degree of freedom approach just like we did with markers, and just like we do with Markerless. You say, well, how accurate can we be at the knee joint. That's the example here this was published by Scott Tashman some time ago. And you can get accuracies down to a bottom millimeter in locating the bone and less than a degree with a repeatability that's really quite high. So you can it's a repeatable measure and it's accurate. What is it a gold standard. Well, sort of sort of a gold standard. What you see here is they have a light from the x-ray source shining towards the race, the receiver, and the light is shining on the subject. That's your entire volume, a bigger volume that requires more extra people don't want that. So here we've got what you see is just the knee. And so then you have to cobble together. This one's cobble together with the rest of the body from markers and just the knee independently. So, can you really verify. Well, you can probably not verify you do the negative. I'll get to that. Now, this seems to come out of the blue but everybody's been waiting for it. If you put markers on the skin. The skin moves relative to the bones. That's just the fact of life, and it moves differently on different parts of the body. It's referred to as soft tissue artifact. And as an aside, there are many, many people get really upset about calling soft tissue movement artifact. For example, all the people in the world trying to build better running bras don't think soft tissue movements and artifact. Nevertheless, when we're talking about rigid objects pose estimation typically used in biomechanics, we think it's an artifact. It's not useful to us. A study was done. What you see here is just a single plane of an x-ray and people can play this again on YouTube we'll see as the markers moving all over the place. So, we're not talking about a teeny tiny little bit of motion. We're talking some serious motion. So here is the study. People ran on a treadmill at the University of Pittsburgh in this case. In front of a bi-planar system you can see the red line showing what they're trying to track. And you see two different views from the two different x-ray images and you see the markers. So we have the relationship between the markers and the bones. And we can track the bones from the x-ray image. And then we're going to introduce something that we're going to call an error. And the error is where is the segment as measured by the markers relative to where the segment is relative to the x-ray. Now, as you can imagine, there are a few things going on here. You have to synchronize your x-ray. This is actually not that trivial to show up. But what you see on the left is you get errors just locating the tibia of 3 millimeters, 4 millimeters. But that's an RMS error. Everybody knows it's probably bigger than that. But sometimes for some people you can get as much as a centimeter of movement. And in this study, these numbers are low. I'll show you another one later. The thigh is a bit worse. That's not surprising. All you have to do is stand there and flex your thigh and you can watch the markers move. So we know they're going to move. And you see upwards of 2 centimeters error in locating the joint. So if we consider x-ray is the gold standard, and it's kind of a flawed gold standard because it's such a small part of the volume. We're in a bit of a conundrum for us. How do you get rid of it? Well, we spent many years trying to get rid of it. One of the very cool ways of getting rid of it is the picture on the left here shows and everybody can go back to this on the video. Look at the original article. It is a lot of markers on the segment. And there are a couple there, if you look closely, that appear to be screwed into the bones. They are. These are pinned right to the bones and there are subjects that will do studies like this. Where they will have bones screwed into, in this case, all the bones of their foot. For the length of the data collection session and then take and unscrew them. I think it is actually very cool. I'd like to mention his name because not only did he have the pin screwed into his bones. He also had three beads implanted in every one of his bones. So after the fact, we get to look at it in x-ray and say, how good are the bone pins? And it turns out then of what they are the bone pins aren't that good. Which is so unfortunate because what a heroic experiment to be a subject in. But let's say we'll go back to our x-ray. Let's say that we've measured the marker movement. Let's say we can describe it. So we have some information on the marker. And many years ago, Thomas Bayes introduced this weird little equation, which I'm not going to describe in a lecture this short, and I'll look it up. But what M.O. Todorov showed us, he published this many years ago, 2007, I think. He said, aha, we can use this information to improve the pose estimation and in effect mitigate the soft tissue artifact on the condition that we can estimate the artifact. Now, it's still a mouthful. So Daniel Walpert, who's very, very clever at making things seem straightforward, gave us this picture on the right here, this tennis court to help us describe what all this is in not so much detail. Anybody wants to go look it up. But this is one lecture. What you see in that image is kind of a yellow blur. That's a tennis ball. You're standing on this side of the court. And your goal is to hit the ball back because that's the game of tennis. In order to do that, it's probably best if you can predict where the ball lands. So from your site, looking at the spin of the ball and way too little time, you might predict that it's going to land in that red ellipse. We'll call that a probability distribution, most likely in the middle. And as you get to the edges, less, less likely it's going to be at the edges. Now, if you actually did a study, you would discover that tennis players wouldn't go to that spot or don't go to that spot. Well, that seems weird. So you're predicting where the ball is going to land, but that's not where you go. That's like, we're going to solve our pose estimation problem for markers or features. And then not use that pose. Well, it's clearly something a little more subtle going on here in the tennis example. Your opponent is actually pretty good. And you've played this opponent many times. And in your experience, your prior history of playing against this opponent. They can't hit along the line. They're good. So your prediction from past performance, some information you have that's not part of this current game or this current shot says, and I really believe it's going to hit the middle of the blue ellipse. I don't think it's going to go into the red. It's going to go in the blue. But do you run there? Well, no, because you've already got this other information. Tonight I've got two different pieces of information. But what Thomas Bay's equation on the left tells us is that there's a, the optimal solution is to multiply the two distributions together. So your confidence in each, it gets multiplying. And you end up in this case going to that pink circle. So back to our pose estimation. The blue is this, well, we know where markers move relative to the bones. We measured it. We did some bi-planar videography. We measured it. We know the markers move relative to the bones. We should use that. So you can build up the results you get from just the pose estimation you recorded. And then you can let it be influenced by what you think should be happening. And that's essentially Bayesian inspired pose estimation because if you want to really read Bayesian pose estimation, you go read some of Todd Pettacchi's work. So I'm calling this Bayesian inspired pose estimation. So what happened when we tried this? As you can see, this is a long time ago. It's hard for me to imagine that the data was collected 20 years ago. It's hard for me to imagine these results are eight years, seven years old. So what you see on the left is us taking a number of subjects, recording the movement. In this case, this is a proof of concept only. We took the lateral knee marker because we know it moves quite a bit. And we just tracked how it moved relative to the shank and the thigh. And then we did this Bayesian inspired pose estimation. And this is pose estimation. We had a few things to think about here. So if you look at that table, it does look kind of small for my screen. Sorry about that. Oh, yes, in many ways we have a solving. We could solve it as six degree freedom. In this case for markers. And we get an error in the shank of 32 millimeters. And then the thigh of 15.6 millimeters. That's pretty big. And if you look at the image there at the top left, you'll see, you know, the white shanks sticking out at the back is not very good. Okay, well, maybe we can constrain it. Very told you about constraints. Well, so let's say the bones can't slide into each other. So we'll take one degree of freedom away. That says these things can't go into each other. Oh my, the shank got worse. The thigh got a little better. But we're still talking about centimeters of error. So then we just pinned it and we said, the knee can't do anything. Well, we get it down now to 13 and 10 millimeters, still quite a bit. But it's really deceiving, because if you pin it, there's no information about what's going on in the joint. But that's the argument that most people have for pinning the knee. They're interfacing the fact that you can't say anything about knee motion for being closer to what you think is a better answer. Now, for this Bayesian inspired one, we took this fairly crude looking regression equation. And we implemented it in a statistical way. And now you're seeing that the errors are down at five and seven millimeters. Well, if one can do that, we'll just model two or three or four. That's a lot of data. And what you see in the right there is in this in that Bayesian sense in that logic, we're throwing everything at this problem. Every prior week and think of and the kitchen sink and that Mary's pizza recipe. Yeah, throw it all in, but do you actually have all that stuff. That's where we got stuck. How much x-ray data would be required to track all the markers that are put on the body for enough variability in the number of subjects to be an accurate representation. Well, we finally had to say, yeah, maybe we can't really mitigate it. This is a mathematical exercise. We can do something we can do. So then you have this pose estimation techniques. And now I'm now I'm actually stealing specific information here. Because I've talked to you about how we do pose talk to you about the number of features that the most critical bit is mathematical observability. I would like to tell you that the critical bit is all these models these priors of soft tissue movement, but we don't have them. So how do we talk about validity. What does it mean. Now, put up a little slide here because I think everybody knows this but it bears repeating. Just a target. Top left, if you're accurate and reliable, reliable in the sense of repeatability, keep getting the same answer. And you'll be in the center of the target. If you're inaccurate and unreliable bottom right. Now you don't get the same thing every time. And it's not near the target. Our problem with pose estimation markers, markerless, inertial sensors, any other body worn sensor for that matter. We can't talk about accuracy. We don't actually have a gold standard. What we can talk about is reliability in the context of repeatability. So I'm going to show you two studies published in the journal by mechanics last year. Just to address this this idea of what's required of us if we wanted if we wanted to do this kind of thing what's required of us. Repeatability. So the study was have some people. These were healthy controls. They come to the lab on, you know, three different days a week apart. And in the case of this study, not in the case of all studies, you have to prescribe or not what they're wearing and what they're doing. In this study, they prescribed the motions and walked back and forth. They didn't really describe what they're wearing said come in whatever you wore to school at day. Got these eight subjects and they come back in. And then we compute from the pose estimation. The joint angles, the joint angles that common to biomechanics, the definitions are common. And what you see on this graph is the mean and standard deviation from each of the data collection sessions. And mostly they lie on top of each other. And you say, that's kind of our definition of repeatable. Now when you get to the bottom right to the top, the left side is flexion extension. You know, almost everybody can measure flexion extension markers, markerless wearables, middle column is abduction abduction still can be done pretty well. Axial rotation. Everybody's working on it. So you can see as you go across to the less reliable joint angles, components of joint angles, then this variability goes up. But how do you talk about it? Think you want to, you want to talk, how do you talk about this variability? You don't, it's not really useful to say, they kind of look like they're on top of each other. So this is one way to look at it. Let's look at the standard deviation across subjects across the gate cycle. And the variability changes depending where in the gate cycle you are. That's not a surprise to anyone. And what you see in the solid line are the three data collection sessions. And the dashed line is one session. So what you can do is just take the ratio of one to the other. The same. Then there's variability ratio would be one. So now we have a number. This is how we can describe repeatability. And we don't have to choose three components of a joint angle. We can choose a different representation. Doesn't really change this story. And that is repeatability is measured as this consistency across trials. And one of the advantages of doing repeatability is there's, there's no argument about how you define your reference frames. We go back to the beginning. Well, you have the same subject. You give them the same model. We're not, they were not describing the difference between models here or difference between techniques or data collection modalities. And kind of comparing apples and apples. But what if you weren't comparing apples and apples? What if you were trying to do something like this? Compare the results of markers and markerless or markers and I am use or I am using markers, whatever. First thing you have to do is make a big decision. How do you define reference frames? Do you impose the same one on both? And if you do, which one do you choose? Because it probably matters. Way, way too many discussions. The decision was, don't even try. Let's define each data collection modality. The way is typically done for that modality. And hope that the reference frames aren't too different. But we only know that at the end, if it turns out the numbers look the same in the end, well, the reference frames are probably the same too. So we'll live with it. And that is the decision. And it's an important one. That you have some idea of these reference frames. And how do you talk about it? And I kind of got away with only presenting joint angles in the repeatability part because, well, they were so similar. That didn't matter how I did it. In this case, how we talk about kinematics, how we talk about the outputs of pose. Adjacent segments actually kind of matter to our interpretation, if the data is different. So on the left side here. What you're seeing is errors. Differences. We don't know that one's right, so differences. And differences of what? These are distances in the top. So that's the difference between the two reference frames. And, and you'll see that it's kind of not great. If you look at the hip, the bottom right, you're seeing up to four centimeters. But I caution you to go back to what we know about soft tissue artifact. And the ability to place markers that we can't do much better than four centimeters. Nevertheless, it's disturbing four centimeters as big. At the ankle and the knee, it's two centimeters and everybody goes, hmm, doesn't really matter how you reference it if you're if you're within that kind of ballpark. And the right side is just the upper extremities treated the same way. But with joint angles, we know that some segments are worse than others. So how do you, how do you talk about it? How do you talk about your post? How to join angles? Or should you talk about the segment relative to the laboratory? And that's what is on the left side here. That is each segment of this lower extremity, the thigh, the shank, and in this case a one segment foot relative to the laboratory. And you'll see quite a bit of consistency across subjects. You'll see, certainly a gain when we get to what was meant by axial rotation, a lot more variability. Now that gives us some clue that the comparison against the lab makes some sense. Is the thigh, the shank, and the foot, are they all okay? The right side is a joint angle. So the thigh relative to the pelvis is the hip. And now you see this bias. Oh, how do we sense that? Well, if we don't think that it's a problem with tracking, that's a problem with the reference frame. So now we've got to go and talk about reference frame differences, decide whether to accept it or not. What I really want to get home is the pose estimation that people have been trying to do my career for 25 years was how to minimize the challenges of reference frame misidentification and soft tissue artifact. And it's not really changed. What I want to capture here is that in pose estimation, there are an awful lot of comparisons going on out there. Everybody wants to believe that if there's a different modality, a different kind of sensor, then we have to compare it against our existing sensor. And the take home message is it's not that easy. We can measure different things really. Not that easy to do the comparison. So let's go through it. Pose estimation, position and orientation of a segment. And usually, of all the segments of the body, that has some anatomical meaning so that when we say flexion extension, we intuitively know what flexion extension means. And based on some feature, some position, I guess could be velocity. There's some feature that we're trying to track that tends to be different for all of the data collection modalities. We assign some meanings some anatomical meaning to it. It's our decision. We, the community want would like to have the same one never quite get there. The tracking just says pose estimation across every frame. And observability says, is it reliable? Could you possibly rely on it? Because it was not mathematically observable. Maybe not. Do we have a gold standard? Not really. Do we have soft tissue artifact? Oh yeah. Do we need it? Theoretically, but no. So can we actually talk about validity? What we can talk about is repeatability and reliability. And perhaps that's the most important thing in this day and age of artificial intelligence and machine learning. If you get the same measure on the same subject doing the same task every time, that's good input for an algorithm. Even if there's some slight bias in terms of accuracy.